TELERILEVAMENTO interno

Snow cover monitoring
with images from digital camera systems
Rosamaria Salvatori 1, Paolo Plini 1, Marco Giusto 1, Mauro Valt 2, Roberto Salzano 1, Mauro
Montagnoli 1, Anselmo Cagnati 2, Giuseppe Crepaz 2 and Daniele Sigismondi 3
1
CNR, Istituto sull’Inquinamento Atmosferico. Via Salaria, km 29,300. 00015 Monterotondo (RM)
2
ARPAV-DRST CVA Via Pradat, 5, 32020 Arabba (BL)
3
Sistemi Video Monitoraggio S.r.l., Strada Provinciale 418, Romito Magra (SP)
E-mail: salvatori@iia.cnr.it
Abstract
Snow cover extension is one of the most important parameters for the study of climate
variations, of hydrological balance and also for the management of touristic activities
in mountain areas. Recently, webcam images collected at daily or even hourly intervals
are used as tools to observe the snow covered areas; those images, properly processed,
can be considered a very important environmental data source. This paper presents the
Snow-noSnow software specifically designed to automatically detect the extension of snow
cover from webcam images. The software was tested on images collected on Alps (ARPAV
webcam network) and on Apennine in a pilot station properly equipped for this project by
CNR-IIA.
Keywords: snow cover monitoring, digital images, software, Alps, Apennines.
Introduction
The seasonal snow cover represents one of the most important land cover class in relation
to environmental studies in mountain areas, especially considering its variation during time.
The snow cover and its extension play a relevant role for the studies on the atmospheric
dynamics and the evolution of climate and also on the analysis and management of water
resources. Moreover, in mountainous areas snow represents a relevant economic resource
for the winter tourism. Considering all these elements, it is clear that monitoring the snow
cover state improves the scientific knowledge on the meteo-climatic phenomena and plays an
important role for the sustainable management on the mountain territory and its resources.
Typically, the snow monitoring is performed through traditional systems like the weather
stations scattered on the territory [Cagnati, 2003], automatic snow height sampling stations
using probes [Cagnati, 1984] or digital cameras [Gook-Hwan et al., 2007], products derived
from satellite images [Casacchia and Salvatori, 2006; Salvatori, 2007; Salzano et al., 2008].
Satellite images like those from NASA-NSIDC (http://nsidc.org/) and NOAA-IMS (http://
www.natice.noaa.gov/ims/) are also used for snow cover daily monitoring at regional scale
[Cianfarra and Valt, 2009; Spisni et al., 2011]. Detailed investigations would require higher
spatial resolution sensors but the revisit time of those sensors is often incompatible with the
Italian Journal of Remote Sensing - 2011, 43 (2): 137-145
doi: 10.5721/ItJRS201143211
137
Salvatori et al.
Snow cover monitoring with digital cameras
snow persistence at soil. Thus images captured by digital cameras become a useful tool at
local scale providing images even when the cloud coverage makes impossible the observation
by satellite sensors. Recently, images taken using digital cameras have been used since they
collect data with a high temporal and spatial resolution at low cost [Hinkler et al. 2002].
Hinkler et al. [2003] have used multispectral images (green, red and near infrared-NIR) taken
using a fixed camera shooting a coastal area in Arctic region (Ny Ålesund, Svalbard), in order
to monitor albedo variations of the snow cover during the melting period. In mountain regions,
Corripio [2004] have used long shot of the Mer de Glace (Mont Blanc), taken with different
field of views (6 megapixel camera) to evaluate the extension of the seasonal snow cover. The
images that have been analysed in the abovementioned papers, have been taken manually and
within a limited period of time. These observing techniques performed using photo cameras in
relation to specific research projects have a limited importance when applied to environmental
monitoring, especially when dealing with the study of climatic variations where extended
data and images- series represent the real added value as shown by Buus-Hinkler et al. [2006]
who have used images captured by a fixed camera and Landsat TM images to monitor the
relationship between snow and vegetation covers in the Arctic environment.
In this perspective the networks of cameras, especially those with a fixed long shot, started
to play a relevant role in an environmental perspective. During the past 10 years, in the
Alpine region many webcams have been installed and their images have been mainly used
for touristic purposes and to a very limited extent for environmental monitoring or in order
to provide support to snow field observations.
When suitably processed these images can be used for scientific purposes, having a good
resolution (at least 800x600x16 million colours) and a very good sampling frequency
(hourly images taken through the whole year). When stored in databases, these images
represent therefore an important source of information for the study of recent climatic
changes, to evaluate the available water resources and to analyse the daily surface evolution
of the snow cover. One of the most important webcam network within the Italian Alps is
the one managed by ARPA Veneto, implemented in 1999 with the financial contribution
of the InterReg II Italia-Austria Programme “AVEN 331010 - Produzione e diffusione
congiunta dei servizi meteorologici a supporto delle attività turistiche delle Alpi orientali”.
The images from the network are being exploited by the personnel charged of weather and
avalanche forecasting in order to perform a visual monitoring of the territory in the Arabba
Avalanche Center (cfr. http://www.arpa.veneto.it/csvdi/svm/webarpav/index.html).
Recently, many webcam have been installed along the Apennines (http://www.
meteoappennino.it/index.php?option=com_webcam&Itemid=86). They are mainly devoted
to have an overlook on the snow cover condition for touristic and recreational use. Cameras
located in the monitoring stations are programmed to take hourly pictures every day. Due to
the high number of collected images there was the need to make available an autonomous
tool to handle such a large image database performing a detailed analysis of them. In order
to obtain quantitative information on the snow cover directly from images, the Italian
National Research Council – Institute for Atmospheric Pollution Research (MonterotondoRome) and ARPA Veneto-Arabba Avalanche Center (CVA) have agreed to collaborate in
order to develop a common project, called Snow-noSnow.
This project has the goal of monitoring areas at different scales, particularly analysing areas
close to the camera and with homogeneous land features in order to obtain information on the
138
Italian Journal of Remote Sensing - 2011, 43 (2): 137-145
snow presence/absence, its persistence and distribution within the examined area. One of the first
results of this project is the development of a dedicated software for the snow cover analysis
through the processing of images taken from webcams and fixed or mobile digital cameras.
The monitoring network
The monitoring equipment installed in two different places in Alps and Apennines is
provided by Sistemi Video Monitoraggio S.r.l. (Romito Magra-SP) and is equipped with
a high resolution digital camera (Canon Powershot SX110 IS, 1/2.3” CCD, 3456x2592
pixel, 6mm objective - 35 mm equivalent focal length of 36mm), a specific hardware for
data logging and transmission, placed into a waterproof case, a power supply unit using AC
or photovoltaic panels with a buffer battery. Data transfer is performed using an intranet
connection with the receiving station located in Arabba through a fax/modem GSM
connection. In the Apennine station, due to the lack of GSM connection, data are stored
locally and automatically copied on a backup hard disk. The system is controlled by the
VM95 software, developed by SVM S.r.l. and Erdmann Video System [Valt, 2002].
The ARPAV station
For the present work the images taken from the camera of the Cima Pradazzo station
(Falcade), (46°21’24”N, 11°49’20”E) have been used. This station is located at 2200 m asl
in the Dolomites. The field view of the camera ranges from a medium shot looking at the
TreValli ski runs to a long shot on the northern part of the Pale di San Martino, the Cime del
Focobon (3054 m) and Monte Mulaz (2906 m). The choice of this field of view is related
to the different kind of snow represented into the image that ranges from untrampled to
groomed snow; the areas in the long shot remain covered by snow for a long time.
The camera is installed close to the snow monitoring station of Cima Pradazzo, equipped
with snow and weather sampling and monitoring instruments, between the others snow
height, internal and surface temperature sensors.
The CNR-IIA Apennine station
In order to make available a series of images of an area in the Apennines, a brand new
experimental station has been conceived and implemented by CNR-IIA in the territory of
the Monti della Laga - Gran Sasso-Monti della Laga National Park. The station is located
along the S.S. 260 “Picente” just after the km 11,500; the camera is mounted on the east
side of a building belonging to the Municipality of Amatrice (Province of Rieti) that make it
available to CNR (42°35.396N, 13°19.787E, 1300 m a.s.l.). Close to the CNR-IIA station, a
manual seasonal weather monitoring station is located in an area called Peschiere (42°60N
13°33E, 1270m asl); it is handled by the Meteomont Service (State Forestry Corps); data
collected only during winter are available on the web. CNR-IIA is going to install an
automatic meteo station close to the camera. The station is equipped with the same digital
camera, hardware and software of the ARPAV stations.
The study area is located on the right bank of the Fosso Cerruglia; it is a slight slope,
characterized by a xeric grassland. The area is easily reachable and scarcely frequented,
it represents a very good point for the observation of the snow cover and the natural
vegetation. In the images the mountain chain of the Monti della Laga comprised between
Monte Gorzano (2458 m) and Cima della Laghetta (2369 m) is distinguishable.
139
Salvatori et al.
The Snow-noSnow Software
Snow cover monitoring with digital cameras
The software Snow-noSnow has been appropriately developed in order to support this
activity; it identifies the extension of the snow cover as shown in the images automatically
taken with a very limited human intervention. The software architecture is based on different
functions representing the preparatory steps of the main processing routine. Snow-noSnow
is handled through an interface where the images to be processed and the ‘mask’ to be
applied for the analysis have to be selected. The mask can be easily prepared using any
drawing software. The areas covered by the mask must be homogeneous; the possibility to
handle a single or multiple masks solves the problem of studying areas with different land
features. The software is able to handle a single or multiple masks. A single mask typically
covers a large part of the image where only the sky and areas at a distance are excluded. The
multiple masks are useful when it is important to analyse different parts of the image with
different surface characteristics. The software has an additional feature, it could select a set
of images to be processed, as specified by date and hour of the capture (Fig. 1).
Each time an image is processed, a specific function is activated; it performs a series of
validations, identifying the images to be rejected due to malfunctions or to bad weather
conditions (heavy snowfall, rainfall or fog). Bad images are identified through a statistical
analysis of RGB values throughout the whole image and are excluded from the processing
steps. The core function of the software is based on a binary snow-no snow classification
algorithm that allows the real identification of snow covered surfaces. The procedure foresees
a statistical analysis of RGB values of all pixels in the mask and allows, using mathematical
criteria, to identify a threshold value to be linked to the snow cover thus allowing its
detection. The classification procedure is based on statistical criteria and does not require
human intervention other than the initial selection of the dataset. The flexibility of the
selection of different parts of the image, through the use of different masks, makes possible
to evaluate different part of territory having different land features. Images are processed
according to the following steps. The jpg image is separated into its RGB components;
for each part of the image inside the predefined mask the Digital Number (DN) frequency
histograms are calculated. A smoothing function is applied to the frequency histogram of
blue component calculated as an average of the 5 nearest points. It is clear that the DN
frequency distribution within the area inside the mask will be different according to the
amount of snow-covered pixels. A preliminary statistical analysis, prior to the definition of
the algorithm was performed analysing around 300 images captured in both areas. It showed
that the blue component corresponding to snow covered areas is always ≥ 127. The histogram
of the values inside the masks shows a bimodal distribution in 90% of cases; when snow is
totally missing the histogram is shifted to low DN values, in case of high snow coverage it is
shifted to highest values (this occurs in all the three components).
To consider the DN frequency distribution, the value of x-axis corresponding to the first local
minimum beyond x≥127 is selected as the threshold value for image classification. When
there are no local minima the DN=127 is selected as the threshold value. When the blue
component obtained from a jpg image is processed, snow always shows higher values also
and mainly in the shaded areas. In fact, in the blue wavelength range, snow cover presents
reflectance values always >0.8 [Salvatori, 2007], which are significantly higher than those
measured on other types of natural surfaces. Therefore during the classification procedure
only the blue component was chosen, thus reducing by one third the processing time.
140
Italian Journal of Remote Sensing - 2011, 43 (2): 137-145
Figure 1 - Screenshot of the Snow-noSnow input interface.
Anyway, applying the same procedure to the three components the same results are
obtained: the unclassified pixels percentage variation is always lower than 1%. Once the
software was compiled, the resulting classifications were compared with those obtained
by photointerpretation and by a segmentation image routine implemented in ENVI. This
last procedure allow us to segment the image into areas of connected pixels based on a
defined range of DN values. For each image the DN values were manually selected from
DN frequency distribution as well as automatically calculated by Snow-noSnow. The
average discrepancy between results automatically obtained from Snow-noSnow and those
obtained running the ENVI segmentation routine for each image, is lower than 0.5%. The
difference is mainly due to how the algorithm identifies the pixels to be connected. Being
the difference negligible, a comparison with the original image was performed based on
visual interpretation; for example considering a 100% snow-covered image, 2-3% of snow141
Salvatori et al.
Snow cover monitoring with digital cameras
covered pixels is not detected by the software due to the presence of shaded snow-covered
associated to the surface roughness.
There are two different modes to use the software in order to calculate the snow cover
surface extension. In the first mode the image is analysed in order to obtain a percentage
value of the snow cover referred to the pixel number. In this case it is sufficient to make
available an image and the mask corresponding to the investigated area. The second mode
requires a more sophisticated elaboration that, taking into account the land features and
the camera field of view, could provide as a result a percentage value correlated to the
actual area expressed in square meters. In this case it is mandatory to use images that can
be corrected according to specific coefficients calculated for the topographic features and
acquisition geometry of the observed area. The following are the preliminary procedures to
adopt for the creation of a weighted mask:
− correction of the image optical deformation. Typically this correction is performed
taking into account the CCD size and the focal length of the camera objective
[Corripio, 2004];
− creation of a digital elevation model (DEM);
− image rectification;
− superimposition of DEM on top of the rectified image;
− definition of the number of pixels for each element of the DEM grid;
− calculation of the ratio between surface unit of each element of the DEM grid and
the number of pixels occurring in the same element;
− provide to each pixel a surface value derived from the size of each element of the
DEM grid. Creation of a ‘weighted mask’ where each pixel into the mask has its
own surface value.
At this point the software is capable to process the image, using the abovementioned routine,
providing a value corresponding to the amount of the snow cover into the study area. The
output of Snow-noSnow is represented by:
− the frequency histograms for each of the three components, where the threshold
used to identify the considered pixels is highlighted;
− the resulting image showing the pixels classified as ‘snow’;
− the value corresponding to the percentage of ‘snow’ pixels compared to the total
amount of pixels into the mask (mode 1) and/or the surface snow covered area
value (mode 2).
When more images are being processed the output file contains threshold value and pixel
percentage values for each of the analysed images. These output features allow to follow
the trend of the snow cover relying on quick processing times being thus useful for people
collecting field measurements.
Results
The analysis of images captured by monitoring stations during the past 3 years in the Alps
and last year in the Apennines shows the capabilities of the Snow-noSnow software to
follow in detail the seasonal and daily evolution of the snow cover. The processing procedure
has been applied to images taken at the CNR-IIA Apennines station where a DEM of the
medium shot area was made available after a topographic survey having the position of the
camera as the reference point. In the Apennines area where snow cover variation could be
142
Italian Journal of Remote Sensing - 2011, 43 (2): 137-145
sudden, the software was particularly effective. The analysis of the hourly images taken at
the CNR-IIA station shows how in less than 24 hours the snow cover could vary from 0 up
to 85% and then back to 0. This shows how meaningful is the variation of snow persistence
at soil in the Apennines. Data that will be collected at the CNR-IIA station in the coming
years would represent a meaningful element for the climate studies at local level. The added
value resulting from this analysis and estimation of the snow cover is particularly relevant
during the deposition and melting phases. During these periods, only punctual data are
collected by traditional instruments and used for different analysis (avalanche forecasts,
estimation of snow water equivalent, dynamics of snow as related to river runoff). As
shown by the analysis of images of Alpine areas, information provided by Snow-noSnow
will compensate the potential estimate errors associated to the snow cover thickness values
measured by a snow gauge. During a melting period, in fact, snow can be discontinuously
distributed on the surface and a snow gauge is only able to detect snow thickness at one
point as occurred in Cima Pradazzo on May 16th 2008 (Fig. 2a).
The spatial analysis performed with Snow-noSnow expressed as a percentage of snow
covered pixels, shows an ablation trend going beyond May 15th. The same occurred during
the ablation phase both in 2009 and 2010 (Fig. 2b, 2c).
Since the software allows to analyse different parts of an image, it is possible to take
advantage of this feature applying it to a mask corresponding to areas that could be
considered an infinite distance away. Knowing the extension value of these areas in square
meters it is possible to estimate the surface of snow covered areas also without a DEM
simply converting the percentage of pixels expressed by the software into a percentage
of square meters into the mask. The images taken in Cima Pradazzo could be analysed to
monitor the snow cover using different masks corresponding to different altitude belts. The
information about the snow coverage in the studied areas were successfully used by snow
technicians to scheduled theirs surveys.
Conclusions and further developments
Images taken from fixed digital cameras could be very useful to monitor snow covers while they
provide data collected continuously, under controlled conditions and also in cloudy conditions.
Taking as true data coming out from photointerpretation, results show that using SnownoSnow only 1% of pixels is misinterpreted.
The results obtained through the use of Snow-noSnow are thus comparable to the one
achieved by photo-interpretation and could be considered as better than the ones obtained
using the image segmentation routine implemented into ENVI. Additionally, Snow-noSnow
operates in a semi-automatic way and has a reduced processing time.
The analysis of this kind of images could represent an useful element to support the interpretation
of remote sensing images, especially those provided by high spatial resolution sensors.
The foreseen improvements would provide unbiased information about some parameters
relates to snow cover expressed until now by subjective evaluations, in order to study the
effects of the wind at high altitudes [Sung-Hyun et al., 2007].
It is foreseen to implement a routine for the estimation of the snow volumes to be used into
the models for the calculation of snow water equivalent.
From the instrumental point of view, it is currently undergoing a testing for the analysis of
snow surface features variation through NIR images.
143
Salvatori et al.
Snow cover monitoring with digital cameras
Figure 2 - Snow cover percentage calculated with SnownoSnow on the image of Cima Pradazzo and snow heigth
measured by the snow gauge during three different yearly
ablation period (a-2008; b-2009, c-2010). The spatial
analysis performed using Snow-noSnow shows the presence
of snow covered areas also when the values of the snow
gauges indicate the absence of snow.
Acknowledgements
The authors would like to acknowledge to the Municipality of Amatrice (Rieti) for its
willingness to host the CNR-IIA equipments.
144
Italian Journal of Remote Sensing - 2011, 43 (2): 137-145
References
Buus-Hinkler J., Hansen B.U., Tamstorf M.P., Pedersen S.B. (2006) - Snow-vegetation
relations in a High Arctic ecosystem: Inter-annual variability inferred from new
monitoring and modeling concepts. Remote Sensing of Environment 105: 237-247.
doi:10.1016/j.rse.2006.06.016
Cagnati A. (1984) - I servizi valanghe regionali e provinciali dell’arco alpino italiano.
Neve e Valanghe, 0: 7-17.
Cagnati A. (2003) - Sistemi di Misura e metodi di osservazione nivometeorologici. AINEVA,
��������
Trento, 186 pp.
Casacchia R., Salvatori R. (2006) - Cartografia delle coperture nevose in ambiente artico.
In: Telerilevamento applicato a cura di Dessena M.A. e Melis M.A. Mako edizioni,
Quartu S.E.
Corripio J.G. (2004) - Snow surface albedo estimating using terrestrial photography. Int. J.
Remote sensing, 25 (24): 5705-5729. doi:10.1080/01431160410001709002
Cianfarra P., Salvini F., Valt M. (2009) - Monitoring the spatio-temporal evolution of the
snow cover in the eastern Alps from MODIS data. Geophysical Research Abstract- Vol.
11, EGU 2009, 8703.
Gook-Hwan K., Sungsoo R., Soon-Geul L. (2007) - Automated Visual Snow-Cover
Measurement Using SRG and VMS. IAENG International Journal of Computer Science,
34:1, IJCS_34_1_22.
Hinkler J., Pedersen S.B., Rasch M., Hansen B.U. (2002) - Automatic snow cover
monitoring at high temporal and spatial resolution, using images taken by a standard
digital camera. International Journal of Remote Sensing 23 (21): 4669-4682. doi:10.10
80/01431160110113881
Hinkler J., Ørbæk J.B., Hansen B.U. (2003) - Detection of spatial, temporal and spectral
surface changes in the Ny-Alesund area 79N, Svalbard, using a low cost multispectral
camera in combination with spectroradiometer measurements. Phys. and Chem. of the
Earth 28: 1229-1239.
Salvatori R. (2007) - Monitoraggio delle coperture nevose con tecniche satellitari per lo
studio dei cambiamenti climatici in aree polari. CNR-Dipartimento
���������������������������������
Terra Ambiente,
“Clima e cambiamenti Climatici, le attività di ricerca del CNR a cura di Carli, Cavaretta,
Fuzzi .pp.307-310 ISBN 978-88-8080-075-0.
Salzano R., Salvatori R., Dominé F. (2008) - Investigation on the relation between physical
and radiometrical properties of snow covers. EARsel eProceedings 7, 1/2008: 68-74.
Spisni A., Tomei F., Pignone S., Muzzi E., Panzacchi A., Antolini G., Villani G., di Lorenzo
M., Foraci R., Bittelli M., Brooks E.S. (2011) - Snow cover analysis in Emilia-Romagna.
Rivista Italiana di Telerilevamento-Italian Journal of Remote Sensing, 43 (1): 59-73.
doi:10.5721/ItJRS20114315
Sung-Hyun L., Gook-Hwan K., Soon-Geul L., Sungsoo R. (2006) - Internet-Based
Visual Snow-Cover Monitoring and Measuring System. Proceedings of SICE-ICASE
International Joint Conference, 3887-3890.
Valt M. (2002) - Le Web Cam di Arabba. Neve e Valanghe, 46: 24-31.
Received 11/02/2011, accepted 15/03/2011
145